Method and apparatus for optical inspection of containers in beverage handling systems

Abstract

The present invention relates to a method (100) for optically inspecting containers in a beverage processing system, wherein the containers are transported (101) using a conveyor as a container quality flow and captured as camera images (102) by an inspection unit arranged in the beverage processing system, and wherein the camera images are inspected for defects by a first evaluation unit using conventional image processing methods (103), wherein camera images of containers with defects are classified as defective images and the defects are assigned as defect labels accordingly to the defective images (104), wherein camera images of containers considered to be of good quality are classified as defect-free images (105), the defective images, the defect labels and the defect-free images are compiled into a specified training dataset (106), and wherein a second evaluation unit (108) is trained on-site using image processing methods based on artificial intelligence using the specified training dataset.

Description

Technical Field

[0001] This invention relates to a method and apparatus for optically inspecting containers in a beverage processing system. Background Technology

[0002] Typically, such methods and equipment are used to inspect containers, for example, to check for contaminants and / or defects. For this purpose, containers are transported by a conveyor as a container quality stream and their images are captured as camera images by an inspection unit arranged within the beverage handling system. The camera images are then inspected for defects by a first evaluation unit using conventional image processing methods. Once a defect is detected, such as contaminants or defects within the container, the container is cleaned or recycled.

[0003] For example, such methods and apparatus for optical inspection of containers are used to inspect the sidewalls, base, and / or fill level of empty or filled containers. It is also conceivable that such methods and apparatus for optical inspection of containers in beverage handling systems are used to detect processing defects during processing within the beverage handling system. For example, whether a container has fallen or is forming a blockage during transportation.

[0004] DE 10 2014 216188A1 discloses an optical inspection method and optical inspection apparatus for containers.

[0005] Typically, image processing methods such as transformation operations, point operations, neighborhood operations, filter operations, histogram operations, threshold operations, brightness operations, and / or contrast operations are used to evaluate camera images, thereby directly detecting image regions with contaminants and / or defects in the camera images.

[0006] The downside here is that the image processing methods used in conventional operations usually require experts to set up for the different types of containers that the beverage handling system is dealing with, which is time-consuming and expensive.

[0007] In addition, artificial intelligence-based image processing methods (e.g., neural networks, especially deep neural networks) are increasingly being used to evaluate camera images.

[0008] The drawback is that such neural or deep neural networks must be trained on a training dataset containing 1,000-10,000 or even more labeled images, each labeled with, for example, blemishes. Labeling is typically done by image processing experts and / or trained staff at beverage processing system manufacturers. This is also time-consuming and expensive. Summary of the Invention

[0009] Therefore, the objective of this invention is to provide a method and apparatus for optically inspecting containers in a beverage processing system, wherein the image processing method is less time-consuming and less costly to set up.

[0010] To address this problem, the present invention provides a method for optically inspecting containers in a beverage processing system, having the features of the present invention.

[0011] By classifying camera images of defective containers as defective images, and correspondingly associating defects with defective images as defect markers, and classifying camera images containing containers found to be good as defect-free images, a large number of defective and defect-free images can be provided based on conventionally operating image processing methods. This can be accomplished, for example, in beverage processing systems with container types, for which conventionally operating image processing methods have been set up and therefore work particularly reliably. Subsequently, a specified training dataset is compiled from the defective images, defect markers, and defect-free images, and this training dataset is used to train a second evaluation unit in the field using an image processing method based on artificial intelligence. Therefore, the specified training dataset can be provided automatically to the greatest extent possible, making the method particularly time-efficient and thus cost-effective.

[0012] Optical inspection methods can be performed before or after the container manufacturing, cleaning, filling, and / or sealing processes, or can be assigned to these processes. For example, the method can be used in full-bottle or empty-bottle inspection machines that include inspection units.

[0013] Containers can be designed to hold beverages, food products, hygiene products, pastes, chemical, biological, and / or pharmaceutical products. Containers can be in the form of bottles, particularly plastic or glass bottles. Specifically, plastic bottles can be PET, PEN, HD-PE, or PP bottles. Containers can also be prefabricated components used to manufacture plastic bottles. Similarly, containers can be biodegradable containers or bottles, primarily composed of renewable raw materials such as sugarcane, wheat, or corn. Containers can have closures, such as crown caps, screw caps, tear-off caps, etc. Likewise, containers can exist as empty containers, preferably without closures.

[0014] It is conceivable that this method could be used to inspect the sidewalls, base, opening, contents, and / or fill level of containers. Defects can be contaminants, such as foreign matter, product residue, label residue, and / or similar substances. Defects can also be flaws, such as damage to the container, particularly cracks and / or broken glass. It is also conceivable that this could be a case of defectively produced material points, such as localized thinning and / or thickening of the material. Furthermore, it is conceivable that this method could be used to inspect returned recyclable containers and / or monitor the transport of containers as a container quality flow and / or monitor the handling of containers in beverage processing systems, for example, detecting containers that have fallen on a conveyor or detecting jamming.

[0015] Containers can be transported to the inspection unit as a container mass flow (preferably a single-channel container mass flow) via a conveyor. However, multi-channel container mass flows are also conceivable. The conveyor may include a turntable and / or a linear conveyor. For example, it is also conceivable that the conveyor may include a conveyor belt on which containers are transported in an upright position to the inspection area of ​​the inspection unit. It is also conceivable that a receiving element receives one or more containers during transport. Containers may also be transported while, for example, the bottom of the container is illuminated by an illumination device and a camera inspects the bottom through the container opening, held by a lateral belt.

[0016] The inspection unit can be configured as an optical inspection unit, particularly with an illumination device and / or a camera to transmit and / or illuminate the container. In the illumination device, light can be generated using at least one light source, for example, a light bulb, a fluorescent tube, and / or at least one LED, to backlight the luminescent surface. Preferably, the light can be generated using an LED matrix and emitted in the direction of the luminescent surface. The luminescent surface can be formed to be larger than the camera's field of view of the container. Similarly, it is conceivable that the luminescent surface only illuminates a portion of the camera's field of view of the container. The luminescent surface can emit light partially or completely diffusely. Preferably, the luminescent surface can include a diffuser plate that diffuses the light from the at least one light source two-dimensionally toward the camera. It is conceivable that the light is generated by the illumination device, then illuminates the container through and / or reflects off it, and is then detected by the camera.

[0017] Using a lens and an image sensor, the camera can capture a portion of a container (specifically, one or several containers) and optionally, light transmitted or reflected from it. The image sensor can be, for example, a CMOS or CCD sensor. It is conceivable that the camera transmits the camera images to a first and / or second evaluation unit via a data interface.

[0018] Each container can be captured in a camera image from one or more image perspectives. The camera can detect the polarization, intensity, color, and / or phase properties of light for each pixel of the camera image.

[0019] The first and / or second evaluation units may process camera images using a signal processor and / or a CPU (Central Processing Unit) and / or a GPU (Graphics Processing Unit) and / or a TPU (Tensor Processing Unit) and / or a VPU (Vision Processing Unit). It is also conceivable that the first and / or second evaluation units include memory units, one or more data interfaces, such as a network interface, a display unit, and / or an input unit. Conventional image processing methods and / or AI-based image processing methods may be implemented in the first and / or second evaluation units (as a computer program product), particularly in the corresponding memory units.

[0020] "Conventional image processing methods" here can mean that conventional image processing methods are not based on artificial intelligence. Specifically, this might mean that conventional image processing methods do not include procedural steps using neural networks, particularly deep neural networks. It is also conceivable that conventional image processing methods evaluate camera images using transformations, points, neighborhoods, filters, histograms, thresholding, brightness, and / or contrast operations in order to directly detect image regions in camera images that include defects.

[0021] "Defect container" can refer to a container that includes defects. "Defect marker" here can refer to a list of defect images that provides an assigned defect description. Similarly, "defect marker" can indicate that a defect marker is entered into the metadata of the corresponding defect image. More generally, "defect marker" can refer to any indicator that a defect image shows a defect. Likewise, "defect marker" might mean the defect coordinates of a defect in a defect image, such as the coordinates of a contaminant. The term "container found to be good" can refer to a container without defects or with tolerable deviations.

[0022] The term "image processing method based on artificial intelligence operations" here can mean that it includes at least one methodological step with a neural network, particularly a deep neural network. For example, an image processing method based on artificial intelligence operations can be a so-called convolutional neural network with at least one convolutional layer and a pooling layer.

[0023] "On-site" can refer to the second evaluation unit being trained on-site or locally by the operator of the beverage processing system using a training dataset, particularly on-site training at the beverage processing system.

[0024] It is conceivable that in the first step, a second evaluation unit utilizing an AI-based image processing method is trained at the manufacturer of the beverage processing system using a general training dataset. Then, in the second step, the second evaluation unit is trained in-situ using a specified training dataset. In this way, the second evaluation unit can be first trained using data available at the manufacturer of the beverage processing system, and then further trained in-situ within the beverage processing system. Therefore, a particularly broad range of data can be obtained for training. The manufacturer of the second evaluation unit can also refer to the manufacturer of the beverage processing system, the inspection unit, and / or the first evaluation unit. The "general training dataset" can represent a compilation of defective images, defect markings, and / or defect-free images of container types known to the manufacturer of the second evaluation unit. Conversely, the specified training dataset can be defective images, defect markings, and defect-free containers of container types known to the operator of the beverage processing system.

[0025] It is conceivable that the training of the second evaluation unit is performed with a lower priority than the inspection unit's capture of the container and / or the first evaluation unit's inspection of camera images, so as to utilize unused resources of the computer system during the inspection period. Therefore, training can be performed during the container inspection without affecting the acquisition of camera images and / or the acquisition of resources used during the inspection of the first evaluation unit. Thus, the computer system can be used particularly efficiently. The computer system can be a machine controller or a PC. The term "lower priority" can refer to the processing priority of the computer system.

[0026] It is conceivable that the recognition performance of the second evaluation unit is determined based on the validation dataset. If the recognition performance exceeds a predetermined threshold, the second evaluation unit uses an AI-based image processing method to check for defects in the camera images. Consequently, the system may automatically switch from checking by the first evaluation unit to checking by the second evaluation unit when the predetermined threshold is reached. The inspection dataset may contain further defective images, defect markers, and / or defect-free images not included in the specified training dataset and / or the general training dataset. Therefore, it can be verified that the AI-based image processing method operates reliably even for containers not checked against the general and / or specified training datasets.

[0027] It is conceivable that the camera images are examined by the second evaluation unit instead of the first evaluation unit. In this way, resources can be used particularly efficiently for the second evaluation unit.

[0028] Alternatively, it is conceivable that camera images are further examined by a first evaluation unit operating in parallel with the second evaluation unit to detect defects that are still unknown to AI-based image processing methods using conventional image processing techniques. This makes it possible to further identify unknown defects and improve the reliability of the evaluation. "Defects that are still unknown to AI-based image processing methods" can refer to defects that have not been classified and compiled into defective images and / or assigned defect labels on a specified training dataset. The evaluation sensitivity of the first evaluation unit can be reduced or set to default parameters to prevent false rejections.

[0029] A camera image of a container with unknown defects can be classified as a further defect image, and the unknown defects can be assigned as further defect labels to the further defect images accordingly. The further defect images and further defect labels can be compiled into a further specified training dataset, and a second evaluation unit can be trained in-situ using an AI-based image processing method with the further specified training dataset. In this way, the recognition performance of the second evaluation unit using the AI-based image processing method can be further improved. It is conceivable that the parallel inspection of camera images by the first and second evaluation units, the classification of camera images of containers with unknown defects, the compilation of the further specified training dataset, and the further training of the AI-based image processing method are all performed iteratively in-situ, specifically until the further recognition performance determined for the second evaluation unit exceeds a predetermined further threshold.

[0030] It is conceivable that defective containers are sorted out of the container stream. This allows defective containers to be excluded from further processing steps and recycled or disposed of.

[0031] Furthermore, in order to solve the above objectives, the present invention provides an apparatus for optically inspecting containers in a beverage processing system, having the features of the present invention.

[0032] The device can be configured to perform the methods of the present invention. In particular, the device can be modified to include, individually or in any combination, the features described above regarding the methods of the present invention.

[0033] By configuring a classification unit to classify camera images of flawed containers as flawed images and assign the flaws accordingly as flaw labels to the flawed images, and to classify camera images of containers found to be good as flawless images, a large number of flawed and flawless images can be provided based on conventionally operating image processing methods. This can be accomplished, for example, in a beverage processing system with container types, for which conventionally operating image processing methods have been set up and are therefore particularly reliable. Since the classification unit is configured to compile flawed images, flaw labels, and flawless images into a specified training dataset, the specified training dataset can be provided for training a second evaluation unit as automatically as possible, meaning that this method is particularly time-efficient and therefore cost-effective.

[0034] The second evaluation unit can be configured to check for defects in the camera image using an AI-based image processing method. Consequently, when a predetermined threshold is exceeded, the system may automatically switch from the first evaluation unit to the second evaluation unit, which then checks for defects in the camera image using the AI-based image processing method. However, it is also conceivable that the first and second evaluation units check the camera image in parallel to increase the overall recognition rate.

[0035] The equipment can be deployed within a beverage processing system. Therefore, the second evaluation unit is trained on-site at the beverage processing system using an AI-based image processing method. Thus, "on-site" can refer to training the second evaluation unit on-site at the beverage manufacturer using a training dataset.

[0036] The device may include a computer system having a first evaluation unit and a second evaluation unit. Therefore, the first evaluation unit and the second evaluation unit can be implemented as a computer program product. It is also conceivable that the computer system includes a classification unit. Thus, the classification unit can also be implemented as a computer program product. The computer system may include a signal processor and / or a CPU (Central Processing Unit) and / or a GPU (Graphics Processing Unit) and / or a TPU (Tensor Processing Unit) and / or a VPU (Visual Processing Unit). It is also conceivable that the computer system includes a memory unit, one or more data interfaces, a network interface, a display unit, and / or an input unit.

[0037] The inspection unit, the first evaluation unit, the second evaluation unit, and / or the classification unit can be interconnected via a digital data link, specifically for transmitting camera images, images with defects, defect markers, images without defects, a specified training dataset, and / or a general training dataset. Attached Figure Description

[0038] Further features and advantages of the invention will be explained in more detail below with reference to the embodiments shown in the accompanying drawings. In the drawings:

[0039] Figure 1 A top view of an embodiment of a device according to the invention for optically inspecting containers in a beverage handling system is shown; and

[0040] Figures 2A to 2B A flowchart according to an embodiment of the present invention is shown, illustrating a method for optically inspecting containers in a beverage processing system. Detailed Implementation

[0041] exist Figure 1 The image shows a top view of an embodiment of a device 1 for optically inspecting a container 2 in a beverage processing system A according to the present invention. Device 1 is configured to perform the functions described below. Figures 2A to 2B Method 100.

[0042] Clearly, container 2 is first transferred via inlet star wheel 9 to filler 6, where it is filled with flowable product. Filler 6 includes, for example, a turntable (not shown) with filling elements arranged thereon, through which container 2 is filled with flowable product during transport. Subsequently, container 2 is transferred via intermediate star wheel 10 to sealing machine 7, where it is provided with a closure, such as a cork, crown stopper, or screw cap. This protects the free-flowing product in container 2 from environmental influences and prevents leakage from container 2.

[0043] Subsequently, container 2 is transferred to conveyor 3 via discharge star wheel 11, and conveyor 3 transports container 2 as a container mass flow to inspection unit 4. The conveyor here is designed, for example, as a conveyor belt that transports container 2 in an upright position. Inspection unit 4, arranged thereon, includes lighting device 42 and camera 41, by which camera 41 captures container 2 in transmitted light. For example, lighting device 42 has a diffuse light disk that uses multiple LEDs to backlight and thus forms an illuminated image background for container 2 as seen from camera 41. Camera 41 then captures container 2 as a camera image, which is forwarded as a digital signal to computer system 5.

[0044] Alternatively or additionally, it is conceivable that container 2 is captured in reflected light by another lighting device.

[0045] Furthermore, a computer system 5 is shown to have a first evaluation unit 51, a second evaluation unit 52, and a classification unit 53. For example, the computer system 5 includes a CPU, a memory unit, an input unit, an output unit, and a network interface. Therefore, the first evaluation unit 51, the second evaluation unit 52, and the classification unit 53 are implemented as a computer program product in the computer system 5.

[0046] The first evaluation unit 51 is configured to use conventional image processing methods to check for defects in the camera image, such as checking for filler levels and / or contaminants.

[0047] Furthermore, classification unit 53 is configured to classify camera images containing defective containers as defective images, assign defects accordingly to defective images as defect labels, and classify camera images containing containers that are found to be good as defect-free images. Additionally, classification unit 53 is configured to compile defective images, defect labels, and defect-free images into a specified training dataset.

[0048] The second evaluation unit 52 is configured to perform an artificial intelligence-based image processing method and train it in-situ using the specified training dataset.

[0049] During the inspection, the first evaluation unit 51 first captures camera images of container 2 and classifies the camera images using classification unit 53 so that a specified training dataset can be compiled from them. Subsequently, the second evaluation unit 52 is then trained live in the beverage processing system A using the specified training dataset.

[0050] The inspection can then be performed alternatively or additionally with the assistance of the second evaluation unit 52. Thus, the second evaluation unit 52 is configured to inspect the camera image for defects using an artificial intelligence-based image processing method.

[0051] Unblemished containers 2 are then fed to further processing steps, such as to a stacker crane. Conversely, defective containers are diverted from the container mass stream by a diverter and then recycled or disposed of.

[0052] The following is for reference. Figures 2A to 2B The operation of the first evaluation unit 51, the second evaluation unit 52, and the classification unit 53 is described in more detail.

[0053] Figures 2A to 2B A flowchart illustrating an embodiment of the present invention is shown, illustrating a method 100 for optically inspecting a container 2 in a beverage handling system A. Reference is made above only as an example. Figure 1 The device 1 for optical inspection described herein is a method 100.

[0054] First, in step 101, container 2 is transported by conveyor 3 as a container mass flow. This is accomplished, for example, by means of a conveyor belt or turntable.

[0055] In subsequent step 102, container 2 is captured as a camera image by inspection unit 4 arranged in the beverage processing system. During this process, container 2 is, for example, illuminated by illumination unit 42 and captured as a camera image in the transmitted light by camera 41. However, it is also conceivable that container 2 may alternatively be illuminated in incident light and captured as a camera image by camera 41.

[0056] In step 103, the first evaluation unit 51 then checks the camera image for defects using conventional image processing methods. Here, "conventional image processing methods" refers to methods without artificial intelligence, such as evaluating the camera image using filters and thresholding operations to detect defects, such as defective filler levels and / or contaminants. For this purpose, for example, an edge filter is used to filter out surfaces of liquid products from the camera image.

[0057] Subsequently, in step 104, the camera image containing the defective container 2 is then classified as a defective image, and the defect is assigned as a defect marker accordingly. For example, the fill level and / or defect indicator can be input as a defect marker into the defective image, specifically in its metadata.

[0058] Furthermore, in step 105, the classification unit 53 classifies camera images containing containers 2 that are found to be in good condition as flawless images.

[0059] Therefore, the flawed images, flawed markers, and flawless images are then compiled into the specified training dataset (step 106).

[0060] For example, classification unit 53 is used to perform these steps 104-106.

[0061] Then, in step 108, the second evaluation unit 52 performs on-site training using an artificial intelligence-based image processing method on a specified training dataset. For example, the deep neural network of the second evaluation unit is trained using the specified training dataset.

[0062] As shown in step 107, it is also conceivable that the second evaluation unit 52, utilizing the AI-based image processing method, is pre-trained at the beverage processing system manufacturer using a general training dataset available therein. Consequently, images of container types and defects already known at the beverage processing system manufacturer can be imported to pre-train the AI-based image processing method, which is then trained in-situ on the beverage processing system A using a specified training dataset. As a result, the AI-based image processing method is further trained on containers of the specified types present at the beverage processing system.

[0063] Additionally, in step 109, the recognition performance of the second evaluation unit 52 is determined based on a validation dataset. For this purpose, the validation dataset may include additional flawed images, flaw markers, and flawless images that are not present in either the general training dataset or the specified training dataset. Therefore, it can be determined whether the second evaluation unit operates reliably.

[0064] If the recognition performance exceeds a predetermined threshold in step 110, the second evaluation unit 52 may use an artificial intelligence-based image processing method to check for defects in the camera image in step 111. This may be done alternatively or in addition to the evaluation by the first evaluation unit 51.

[0065] By alternatively having the evaluation performed by a second evaluation unit, reliable recognition performance can be ensured through artificial intelligence, without the need for time-consuming parameterization of classical image processing methods by experts.

[0066] On the other hand, if the camera images are examined in parallel by the first evaluation unit 51 and the second evaluation unit 52, it is possible to detect defects that are still unknown to the AI-based image processing method by using conventional image processing methods. This makes it possible to improve the reliability of the evaluation. Therefore, the evaluation sensitivity of the first evaluation unit 51 can be reduced or set to a standard parameter to prevent false rejections. It is conceivable that camera images representing containers 2 with unknown defects are classified as further defect images, and the unknown defects are correspondingly assigned as further defect labels to the further defect images, wherein the further defect images and further defect labels are compiled into a further specified training dataset, and wherein the second evaluation unit 52 is further trained in-situ using the AI-based image processing method using the further specified training dataset. As a result, the recognition performance of the second evaluation unit 52 utilizing the AI-based image processing method can be further improved.

[0067] On the other hand, if the recognition performance in step 110 does not exceed a predetermined threshold, the camera images from step 110 are still inspected for defects by the first evaluation unit 51 using image processing methods operating on a conventional intelligence basis. In this case, the classification unit 53 can further classify the camera images according to steps 104 and 105 and add further defective images, defective labels, and defect-free images to a specified training dataset or generate another specified training dataset. This can then be used to further train the second evaluation unit 52 according to step 108 until the recognition performance in step 110 exceeds the predetermined threshold.

[0068] It is also conceivable that the training of the second evaluation unit 52 in step 108 be performed with a lower priority than the capture of the container by the inspection unit 4 in step 102 and / or the inspection of the camera image by the first evaluation unit 51 in step 103, so as to utilize the unused resources of the computer system 5 during the inspection.

[0069] By classifying camera images of defective containers 2 as flawed images and assigning defects as flaw labels accordingly, and classifying camera images of containers 2 that are found to be good as flawless images, a large number of flawed and flawless images can be provided based on conventionally operating image processing methods. This can be accomplished, for example, in a beverage processing system A with container types, for which conventionally operating image processing methods have been set up and are therefore particularly reliable. Subsequently, a specified training dataset is compiled from the flawed images, flaw labels, and flawless images, and the second evaluation unit 52 is thus trained in-situ using an image processing method based on artificial intelligence. Therefore, the specified training dataset can be provided automatically to the greatest extent possible, thereby allowing method 100 to operate in a particularly time-efficient and therefore cost-effective manner.

[0070] It should be understood that the features mentioned in the above embodiments are not limited to these feature combinations, but can also be implemented individually or in any other feature combination.

Claims

1. A method (100) for optically inspecting a container in a beverage handling system, wherein, The container is transported by a transport aircraft as a container mass flow (101), and the container is captured as a camera image by an inspection unit arranged in the beverage processing system (102), wherein the camera image is inspected for defects by a first evaluation unit using a conventional image processing method (103), wherein the conventional image processing method does not include a process step using a neural network. Its features are, Camera images with defective containers are classified as defective images, and the defects are assigned as defect labels to the defective images accordingly (104). Camera images with containers that were found to be in good condition were classified as flawless images (105). The flawed image, the flaw marker, and the flawless image are compiled into a specified training dataset (106), and The second evaluation unit (108) using the specified training dataset is trained in the field using an image processing method based on artificial intelligence operations. In the first step, the second evaluation unit (107) using an image processing method based on artificial intelligence operation is trained at the manufacturer of the second evaluation unit using a general training dataset. Then, in the second step, the second evaluation unit (108) using the specified training dataset is trained on-site using the image processing method based on artificial intelligence operation.

2. The method (100) according to claim 1, wherein, The training (108) of the second evaluation unit is performed with a lower priority than the inspection unit capturing the container (102) and / or the first evaluation unit inspecting the camera image (103), so as to utilize unused resources of the computer system during inspection.

3. The method (100) according to claim 1 or 2, wherein, The recognition performance of the second evaluation unit is determined based on the verification dataset (109), and if the recognition performance exceeds a predetermined threshold (110), the second evaluation unit uses an image processing method based on artificial intelligence operations to check for defects in the camera image (111).

4. The method (100) according to claim 3, wherein, The camera images are examined by the second evaluation unit, not the first evaluation unit.

5. The method (100) according to claim 3, wherein, The camera images are further examined by the first evaluation unit, which operates in parallel with the second evaluation unit, in order to capture flaws that are still unknown to AI-based image processing methods by means of conventional image processing methods.

6. The method (100) according to claim 5, wherein, Camera images of containers with unknown defects are classified as further defect images, and the unknown defects are correspondingly assigned as further defect tags to the further defect images. The further defect images and the further defect markers are compiled into a further specified training dataset, and The further specified training dataset is used to train the second evaluation unit on-site using an image processing method based on artificial intelligence operations.

7. The method (100) according to claim 5, wherein, The evaluation sensitivity of the first evaluation unit is reduced or set to a default parameter to prevent false rejection.

8. The method (100) according to claim 1 or 2, wherein, The defective containers are sorted out from the container quality stream.

9. An apparatus (1) for optically inspecting a container (2) in a beverage handling system (A), the apparatus (1) being used to perform the method according to any one of claims 1 to 8, the apparatus comprising: The transport aircraft (3) is used to transport the container (2) as a container mass flow. Inspection unit (4), which is arranged in the beverage processing system (1) to capture the container (2) as a camera image, and A first evaluation unit (51) is configured to examine defects in the camera image using conventional image processing methods, wherein the conventional image processing methods do not include a process step using a neural network. Its features are, The classification unit (53) is configured to classify camera images with flawed containers as flawed images and assign the flaws as flaw labels accordingly to the flawed images, classify camera images with containers that are found to be good as flawless images, and compile the flawed images, the flaw labels, and the flawless images into a specified training dataset. The second evaluation unit (52) is configured to perform an image processing method based on artificial intelligence operations and to train the image processing method based on artificial intelligence operations in the field using the specified training dataset.

10. The device (1) according to claim 9, wherein, The second evaluation unit (52) is configured to use an image processing method based on artificial intelligence operations to check for defects in the camera image.

11. The device (1) according to claim 9 or 10, wherein, The device (1) is arranged in the beverage processing system (A).

12. The device (1) according to claim 9 or 10, wherein, The device (1) includes a computer system (5), which includes the first evaluation unit (51) and the second evaluation unit (52).

Images